Free word lists (SCOWL, ENABLE, TWL) provide validated spellings—176K to 267K words depending on the list. Academic resources like WordNet and Wiktionary add definitions and taxonomies. Commercial publishers like Oxford provide ~600K synonyms. None include difficulty rankings, content filters, or weighted semantic connections.
A 4×4 categorization puzzle takes an expert about two hours. A 6×8 takes half a day. The workload scales nonlinearly: every additional category multiplies cross-checks against every other. Wordle’s curated answer list of ~2,300 words started recycling in February 2026.
Ask an LLM for 50 word-association puzzles and the first five look great. By puzzle 20, the same animals, the same colors, the same “things that are round.” Each prompt is solid on its own—but there’s no memory across the set.
Generating thousands of non-repeating puzzles requires two things simultaneously: deep language data and global state across the full vocabulary—knowing which associations have been used, which senses are underrepresented, which difficulty bands need more content.
It’s not an accident that the publicly available alternatives are thin. The legal and economic dynamics of language data create a market where the best work stays invisible. Here’s why the best language data is not on GitHub.